Skip to content

Release Notes — v0.7.11

MADOCA regression tests: build-independent upstream parity + absolute accuracy

Release date: 2026-09-29 Type: Test infrastructure — no positioning change (no engine, decoder or numerical code changed) Branch: release/v0.7.11

Overview

v0.7.11 is not a feature release. It fixes the MADOCA regression tests so the full suite passes on every supported build, and gives every MADOCA case the same two-tier check the CLAS tests already have: upstream parity (port fidelity) plus absolute accuracy against independent truth, which stays meaningful as MRTKLIB diverges from or improves on upstream.

madocalib_pppar_ion_check passed when its 5 mm tolerance was introduced in v0.3.1 (0.25 cm), but from v0.6.6 onward it failed on every build that links LAPACK (Accelerate on macOS), and release notes carried it as a known "environment-only" failure. CI stayed green only because it installs no BLAS and CMake then switched the same test to an 8× looser tolerance. With this release the suite is 125 / 125 on macOS and the strict gate runs in CI.

Root cause (#302)

Measured on macOS with upstream MADOCALIB and MRTKLIB each built against internal LU, Accelerate and OpenBLAS:

  • Not nondeterminism. Every build is bit-identical run-to-run.
  • Not an MRTKLIB regression. Against a same-backend upstream build MRTKLIB differs by 0.236 cm, matching what was recorded when the reference was introduced.
  • A stale, backend-specific reference. The pppar / pppar_ion references came from an upstream Accelerate build that no longer reproduces on the current toolchain (today's upstream Accelerate build is 1.62 cm away, one fix/float flip at 00:29).
  • A metric that cannot tolerate any backend change. This one-hour dataset has several epochs where PPP-AR sits at the ratio-test margin; a ULP-level difference flips fix ↔ float there, and one flip is 17–24 cm. Pinning both sides to OpenBLAS still fails (2.2 cm), because port and upstream are different compiled code. At 00:29 the fixed solution is closer to SINEX truth than the reference's float (2D 4.1 vs 9.6 cm), so the old gate penalised the better answer.

Changes

  1. References regenerated from upstream as shipped — internal LU, no -DLAPACK; platform-independent and what CI builds. The generator was passing MRTKLIB TOML to upstream rnx2rtkp (which reads only key = value .conf); upstream-format configs are restored in tests/data/madocalib/upstream_conf/, and each regeneration writes reference_provenance.txt (upstream version and commit, backend, flags, compiler, platform). pp / pppar_003 are byte-identical; pppar / pppar_ion moved by 1.51 / 3.78 cm (3D RMS).
  2. Parity tolerances are the same in every build — the LAPACK_FOUND switch is gone.
  3. Integer-fix-rate gate for PPP-AR parity. compare_pos.py counted float PPP (Q=6) as fixed and computed the rate over the ref/test intersection, so neither a fix→float drop nor a lost epoch was visible. PPP-AR checks now count Q=1 only over the reference timeline (a missing test epoch is not fixed), one-sided at −5 %. New options --fix-q and --max-fix-drop; defaults preserve the old behaviour for other callers.
  4. Absolute checks for all four MADOCA cases against a MIZU-only extract of the week-2360 IGS SINEX (2025/03/30–04/05, contains the data epoch; 977 bytes). It replaces the 15 MB week-2383 file, taken five months after the data without velocity propagation (MIZU moved 2.26 cm in between — about half the previously reported PPP-AR error).
  5. CI pins the matrix backend (-DCMAKE_DISABLE_FIND_PACKAGE_LAPACK=TRUE in the regression job) and configure prints Matrix backend: ….

Tolerance changes

Test Before After Measured worst (LU / Accelerate / OpenBLAS)
madocalib_pppar_check 0.008 m (LAPACK) / 0.020 m 0.020 m 1.563 cm
madocalib_pppar_ion_check 0.005 m (LAPACK) / 0.040 m 0.050 m 3.397 cm
PPP-AR fix rate (3 checks) Q ∈ {1,6}, −1 %, intersection Q = 1, −5 %, reference timeline −3.39 %
madocalib_ppp_abs_check — 0.50 m 24.36 cm (2D 95 %)
madocalib_pppar_abs_check 0.100 m (week 2383) 0.06 m (week 2360) 2.66 cm
madocalib_pppar_003_abs_check — 0.08 m 3.65 cm
madocalib_pppar_ion_abs_check — 0.06 m 2.75 cm

The parity loosening is offset by the new fix-rate gate and absolute checks: removing 7 fixed epochs from a pppar_ion output, or turning them into float, passed the old checks (+0.00 %) and fails the new one (−5.93 %).

Verification

  • MADOCA tests (15) pass on macOS internal-LU, Accelerate and OpenBLAS (MRTK_DETERMINISTIC_BLAS) builds.
  • Full suite on macOS (Accelerate): 125 / 125.
  • Linux CI (internal LU, pinned): 120 / 120 (-LE realtime).
  • No positioning code changed.
  • #302 — fixed (PR #338).
  • #300 — multi-dimensional acceptance framework; the CLAS compare_nmea.py fix rate has the same "counts non-RTK-fix as fix" weakness and is left for that work.

Upgrade notes

None for library or CLI users. Contributors: madocalib_pppar_ion_check is no longer an expected failure — treat any MADOCA check failure as real.